Claude Codex Marketplace

A Claude Code plugin marketplace providing multi-AI orchestration tools for planning, implementing, and reviewing code changes.

New to Claude Codex? Check out our Wiki for beginner-friendly guides and tutorials. Available in English and 中文.

Why Multi-AI Review?

Would you deploy code that only one person looked at?

In professional software development, code reviews are mandatory. Google requires review for every change. Microsoft uses multiple review stages. Claude Codex brings this professional standard to AI-assisted development.

Instead of trusting a single AI's output, your code goes through three independent reviews:

Reviewer	What It Catches
Claude Sonnet	Obvious bugs, security basics, code style
Claude Opus	Architectural issues, subtle bugs, edge cases
Codex	Fresh perspective from a different AI

The result: More bugs caught, better security, and production-ready code.

Each reviewer checks for OWASP Top 10 vulnerabilities, proper error handling, and code quality. If Sonnet misses something, Opus or Codex will likely catch it. The loop-until-approved model means code doesn't proceed until all three reviewers give the green light.

Learn more: Why Claude Codex?

Installation

Step 1: Add Marketplace

/plugin marketplace add Z-M-Huang/claude-codex

Step 2: Install Plugin

# Install at user scope (available in all projects) - RECOMMENDED
/plugin install claude-codex@claude-codex --scope user

# OR install at project scope (this project only)
/plugin install claude-codex@claude-codex --scope project

Step 3: Add .task to .gitignore

The plugin automatically creates .task/ in your project directory. Add it to .gitignore:

echo ".task" >> .gitignore

Multi-Project Support: When installed at user scope, you can run the pipeline in multiple projects simultaneously. Each project gets its own isolated .task/ directory.

Usage

After installation, use skills with the plugin namespace:

# Start the full pipeline
/claude-codex:multi-ai Add user authentication with JWT

Available Plugins

claude-codex

Multi-AI orchestration pipeline with hook-based enforcement using Task + Resume architecture. Skills:

Skill	Purpose
multi-ai	Pipeline entry point (orchestrates agents)

Custom Agents:

Agent	Model	Purpose
requirements-gatherer	Opus	Business Analyst + PM hybrid
planner	Opus	Architect + Fullstack hybrid
plan-reviewer	Both	Architecture + Security + QA validation
implementer	Sonnet	Fullstack + TDD + Quality implementation
code-reviewer	Both	Security + Performance + QA validation

Recommended Subscriptions

Service	Subscription	Purpose
Claude Code	MAX 20	Main thread (planning, coding) + Review skills
Codex CLI	Plus	Final reviews (invoked via skill)

How It Works

Quick Start with /multi-ai

/claude-codex:multi-ai Add user authentication with JWT tokens

This command orchestrates specialized agents:

1. Requirements (team-based) - Specialist teammates explore in parallel, lead asks informed questions, requirements-gatherer synthesizes
2. Planning (semi-interactive) - planner agent creates plan, asks user only if needed
3. Plan Reviews (autonomous) - plan-reviewer agents (sonnet, opus) + Codex gate
4. Implementation (autonomous) - implementer agent iterates until tests pass + reviews approve
5. Code Reviews (autonomous) - code-reviewer agents (sonnet, opus) + Codex gate
6. Complete - Reports results

Task-Based Pipeline Enforcement

The pipeline uses task dependencies, hook validation, and agent teams (v1.4.2):

┌────────────────────────────────────────────────────┐
│  SEQUENTIAL REVIEWS (enforced via blockedBy)        │
│  ┌──────────────────────────────────────────────┐  │
│  │ 1. Implement code (implementer agent)        │  │
│  │ 2. Code review (sonnet) - blocked until impl │  │
│  │ 3. Code review (opus) - blocked until sonnet │  │
│  │ 4. Code review (codex) - blocked until opus  │  │
│  │                                              │  │
│  │ IF review needs_changes                      │  │
│  │    → Create fix task + re-review task        │  │
│  │    → Same reviewer validates fixes           │  │
│  │                                              │  │
│  │ IF all reviews approved                      │  │
│  │    → Pipeline complete                       │  │
│  └──────────────────────────────────────────────┘  │
└────────────────────────────────────────────────────┘

Key benefits:

• Hook enforcement - SubagentStop validates acceptance criteria coverage
• Task dependencies - blockedBy prevents skipping reviews
• Same-reviewer validation - Fixes are validated by the reviewer who requested them
• Clarification support - Reviewers can request clarification via needs_clarification status

Marketplace Structure

claude-codex/
├── .claude-plugin/
│   └── marketplace.json          # Marketplace catalog
├── plugins/
│   └── claude-codex/             # Multi-AI plugin
│       ├── .claude-plugin/
│       │   └── plugin.json       # Plugin manifest
│       ├── agents/               # Custom agent definitions
│       │   ├── requirements-gatherer.md
│       │   ├── planner.md
│       │   ├── plan-reviewer.md
│       │   ├── implementer.md
│       │   └── code-reviewer.md
│       ├── skills/               # Pipeline skills
│       │   └── multi-ai/         # Main orchestrator
│       ├── scripts/              # Orchestration scripts
│       ├── hooks/                # Pipeline enforcement hooks
│       ├── docs/                 # Standards and workflow
│       ├── .task.template/       # Task directory template
│       ├── CLAUDE.md
│       └── AGENTS.md
└── README.md

Prerequisites

• Claude Code installed and authenticated
• Codex CLI installed and authenticated (for review-codex skill)
• Bun installed (required by Claude Code, also used for cross-platform JSON processing)

Note: This plugin works on Windows, macOS, and Linux. All shell scripts use Bun for JSON processing instead of jq, ensuring cross-platform compatibility out of the box.

Default Settings

The pipeline uses these hardcoded defaults:

Setting	Default	Description
Plan review loop limit	10	Max iterations for plan reviews
Code review loop limit	15	Max iterations for code reviews
Auto-resolve attempts	3	Retries before pausing on errors

Creating Your Own Plugin

To add a new plugin to this marketplace:

1. Create a directory under plugins/your-plugin-name/
2. Add .claude-plugin/plugin.json with your plugin manifest
3. Add your skills under skills/
4. Update .claude-plugin/marketplace.json to include your plugin

Example plugin.json:

{
  "name": "your-plugin-name",
  "version": "1.0.9",
  "description": "What your plugin does",
  "author": { "name": "Your Name" },
  "skills": "./skills/"
}

Troubleshooting

Plugin not found

# Verify marketplace is added
/plugin marketplace list

# Re-add if needed
/plugin marketplace add Z-M-Huang/claude-codex

Skills not loading

# Check plugin is installed
/plugin list

# Reinstall if needed
/plugin uninstall claude-codex@claude-codex
/plugin install claude-codex@claude-codex --scope user

Validate plugin structure

/plugin validate .

Skill not working with slash command

When invoking skills from external projects, you must use the full namespaced format:

# Correct - use the full namespace
/claude-codex:multi-ai Add user authentication

# Wrong - bare skill name doesn't work from external projects
/multi-ai Add user authentication

The bare skill name (e.g., /multi-ai) only works within the plugin's internal context. When using the plugin from your own project, always prefix with claude-codex:.

Alternative: You can also ask Claude naturally without using slash commands:

"Use the multi-ai pipeline to add user authentication"

Claude will recognize this and invoke the appropriate skill.

Documentation

For detailed guides, visit our Wiki:

English

• Why Claude Codex? - Benefits of multi-AI review
• Getting Started - Step-by-step installation for beginners
• Understanding the Pipeline - How the review process works
• How to Use - Commands and usage examples
• Configuration - Customization options
• Troubleshooting - Common issues and solutions

中文

• 为什么选择 Claude Codex？ - 多 AI 审查的优势
• 快速开始 - 零基础安装指南
• 流水线详解 - 审查流程工作原理
• 使用方法 - 命令和使用示例
• 配置说明 - 自定义选项
• 常见问题 - 问题排查与解决

Related Projects

• claude-codex-gemini - Adds Gemini as a dedicated orchestrator

License

GPL-3.0 with attribution requirement.

If you use, modify, or distribute this software, please provide attribution to:

• Author: Z-M-Huang
• Project: Claude Codex
• Repository: https://github.com/Z-M-Huang/claude-codex

See LICENSE for full terms.